diff --git a/fastNLP/io/data_loader/matching.py b/fastNLP/io/data_loader/matching.py
index 771f2748..ce9c280b 100644
--- a/fastNLP/io/data_loader/matching.py
+++ b/fastNLP/io/data_loader/matching.py
@@ -1,6 +1,6 @@
 import os
 
-from typing import Union, Dict
+from typing import Union, Dict , List
 
 from ...core.const import Const
 from ...core.vocabulary import Vocabulary
@@ -33,7 +33,8 @@ class MatchingLoader(DataSetLoader):
                 to_lower=False, seq_len_type: str=None, bert_tokenizer: str=None,
                 cut_text: int = None, get_index=True, auto_pad_length: int=None,
                 auto_pad_token: str='<pad>', set_input: Union[list, str, bool]=True,
-                set_target: Union[list, str, bool] = True, concat: Union[str, list, bool]=None, ) -> DataInfo:
+                set_target: Union[list, str, bool] = True, concat: Union[str, list, bool]=None, 
+                extra_split: List[str]=['-'], ) -> DataInfo:
         """
         :param paths: str或者Dict[str, str]。如果是str，则为数据集所在的文件夹或者是全路径文件名：如果是文件夹，
             则会从self.paths里面找对应的数据集名称与文件名。如果是Dict，则为数据集名称（如train、dev、test）和
@@ -56,6 +57,7 @@ class MatchingLoader(DataSetLoader):
         :param concat: 是否需要将两个句子拼接起来。如果为False则不会拼接。如果为True则会在两个句子之间插入一个<sep>。
             如果传入一个长度为4的list，则分别表示插在第一句开始前、第一句结束后、第二句开始前、第二句结束后的标识符。如果
             传入字符串 ``bert`` ，则会采用bert的拼接方式，等价于['[CLS]', '[SEP]', '', '[SEP]'].
+        :param extra_split: 额外的分隔符，即除了空格之外的用于分词的字符。
         :return:
         """
         if isinstance(set_input, str):
@@ -89,6 +91,24 @@ class MatchingLoader(DataSetLoader):
                 if Const.TARGET in data_set.get_field_names():
                     data_set.set_target(Const.TARGET)
 
+        if extra_split:
+            for data_name, data_set in data_info.datasets.items():
+                data_set.apply(lambda x: ' '.join(x[Const.INPUTS(0)]), new_field_name=Const.INPUTS(0))
+                data_set.apply(lambda x: ' '.join(x[Const.INPUTS(1)]), new_field_name=Const.INPUTS(1))
+
+                for s in extra_split:
+                    data_set.apply(lambda x: x[Const.INPUTS(0)].replace(s , ' ' + s + ' '), 
+                                    new_field_name=Const.INPUTS(0))
+                    data_set.apply(lambda x: x[Const.INPUTS(0)].replace(s , ' ' + s + ' '), 
+                                    new_field_name=Const.INPUTS(0))
+
+                _filt = lambda x : x
+                data_set.apply(lambda x: list(filter(_filt , x[Const.INPUTS(0)].split(' '))), 
+                                    new_field_name=Const.INPUTS(0), is_input=auto_set_input)
+                data_set.apply(lambda x: list(filter(_filt , x[Const.INPUTS(1)].split(' '))), 
+                                    new_field_name=Const.INPUTS(1), is_input=auto_set_input)
+                _filt = None
+
         if to_lower:
             for data_name, data_set in data_info.datasets.items():
                 data_set.apply(lambda x: [w.lower() for w in x[Const.INPUTS(0)]], new_field_name=Const.INPUTS(0),
diff --git a/reproduction/matching/matching_mwan.py b/reproduction/matching/matching_mwan.py
new file mode 100644
index 00000000..d2d3033f
--- /dev/null
+++ b/reproduction/matching/matching_mwan.py
@@ -0,0 +1,145 @@
+import sys
+
+import os
+import random
+
+import numpy as np
+import torch
+from torch.optim import Adadelta, SGD
+from torch.optim.lr_scheduler import StepLR
+
+from tqdm import tqdm
+
+from fastNLP import CrossEntropyLoss
+from fastNLP import cache_results
+from fastNLP.core import Trainer, Tester, Adam, AccuracyMetric, Const
+from fastNLP.core.predictor import Predictor
+from fastNLP.core.callback import GradientClipCallback, LRScheduler, FitlogCallback
+from fastNLP.modules.encoder.embedding import ElmoEmbedding, StaticEmbedding
+
+from fastNLP.io.data_loader import MNLILoader, QNLILoader, QuoraLoader, SNLILoader, RTELoader
+from model.mwan import MwanModel
+
+import fitlog
+fitlog.debug()
+
+import argparse
+
+
+argument = argparse.ArgumentParser()
+argument.add_argument('--task'         , choices = ['snli', 'rte', 'qnli', 'mnli'],default = 'snli')
+argument.add_argument('--batch-size'   , type = int    ,   default = 128)
+argument.add_argument('--n-epochs'     , type = int    ,   default = 50)
+argument.add_argument('--lr'           , type = float  ,   default = 1)
+argument.add_argument('--testset-name' , type = str    ,   default = 'test')
+argument.add_argument('--devset-name'  , type = str    ,   default = 'dev')
+argument.add_argument('--seed'         , type = int    ,   default = 42)
+argument.add_argument('--hidden-size'  , type = int    ,   default = 150)
+argument.add_argument('--dropout'      , type = float  ,   default = 0.3)
+arg = argument.parse_args()
+
+random.seed(arg.seed)
+np.random.seed(arg.seed)
+torch.manual_seed(arg.seed)
+
+n_gpu = torch.cuda.device_count()
+if n_gpu > 0:
+    torch.cuda.manual_seed_all(arg.seed)
+print (n_gpu)
+
+for k in arg.__dict__:
+    print(k, arg.__dict__[k], type(arg.__dict__[k]))
+
+# load data set
+if arg.task == 'snli':
+    @cache_results(f'snli_mwan.pkl')
+    def read_snli():
+        data_info = SNLILoader().process(
+            paths='path/to/snli/data', to_lower=True, seq_len_type=None, bert_tokenizer=None,
+            get_index=True, concat=False, extra_split=['/','%','-'],
+        )
+        return data_info
+    data_info = read_snli()
+elif arg.task == 'rte':
+    @cache_results(f'rte_mwan.pkl')
+    def read_rte():
+        data_info = RTELoader().process(
+            paths='path/to/rte/data', to_lower=True, seq_len_type=None, bert_tokenizer=None,
+            get_index=True, concat=False, extra_split=['/','%','-'],
+        )
+        return data_info
+    data_info = read_rte()
+elif arg.task == 'qnli':
+    data_info = QNLILoader().process(
+        paths='path/to/qnli/data', to_lower=True, seq_len_type=None, bert_tokenizer=None,
+        get_index=True, concat=False , cut_text=512, extra_split=['/','%','-'],
+    )
+elif arg.task == 'mnli':
+    @cache_results(f'mnli_v0.9_mwan.pkl')
+    def read_mnli():
+        data_info = MNLILoader().process(
+            paths='path/to/mnli/data', to_lower=True, seq_len_type=None, bert_tokenizer=None,
+            get_index=True, concat=False, extra_split=['/','%','-'],
+        )
+        return data_info
+    data_info = read_mnli()
+else:
+    raise RuntimeError(f'NOT support {arg.task} task yet!')
+
+print(data_info)
+print(len(data_info.vocabs['words']))
+
+
+model = MwanModel(
+	num_class = len(data_info.vocabs[Const.TARGET]), 
+	EmbLayer = StaticEmbedding(data_info.vocabs[Const.INPUT], requires_grad=False, normalize=False), 
+	ElmoLayer = None,
+	args_of_imm = {
+		"input_size"	 	: 300 , 
+		"hidden_size" 		: arg.hidden_size , 
+		"dropout" 			: arg.dropout	,
+		"use_allennlp" 		: False , 
+	} , 
+)
+
+
+optimizer = Adadelta(lr=arg.lr, params=model.parameters())
+scheduler = StepLR(optimizer, step_size=10, gamma=0.5)
+
+callbacks = [
+    LRScheduler(scheduler),
+]
+
+if arg.task in ['snli']:
+    callbacks.append(FitlogCallback(data_info.datasets[arg.testset_name], verbose=1))
+elif arg.task == 'mnli':
+    callbacks.append(FitlogCallback({'dev_matched': data_info.datasets['dev_matched'],
+                                     'dev_mismatched': data_info.datasets['dev_mismatched']},
+                                    verbose=1))
+
+trainer = Trainer(
+    train_data       = data_info.datasets['train'], 
+    model            = model,
+    optimizer        = optimizer, 
+    num_workers      = 0,
+    batch_size       = arg.batch_size,
+    n_epochs         = arg.n_epochs, 
+    print_every      = -1,
+    dev_data         = data_info.datasets[arg.devset_name],
+    metrics          = AccuracyMetric(pred = "pred" , target = "target"), 
+    metric_key       = 'acc', 
+    device           = [i for i in range(torch.cuda.device_count())],
+    check_code_level = -1, 
+    callbacks        = callbacks,
+    loss             = CrossEntropyLoss(pred = "pred" , target = "target")
+)
+trainer.train(load_best_model=True)
+
+tester = Tester(
+    data=data_info.datasets[arg.testset_name],
+    model=model,
+    metrics=AccuracyMetric(),
+    batch_size=arg.batch_size,
+    device=[i for i in range(torch.cuda.device_count())],
+)
+tester.test()
diff --git a/reproduction/matching/model/mwan.py b/reproduction/matching/model/mwan.py
new file mode 100644
index 00000000..7ca6df3b
--- /dev/null
+++ b/reproduction/matching/model/mwan.py
@@ -0,0 +1,455 @@
+import torch as tc
+import torch.nn as nn
+import torch.nn.functional as F
+import sys
+import os
+import math
+from fastNLP.core.const import Const
+
+class RNNModel(nn.Module):
+    def __init__(self, input_size, hidden_size, num_layers, bidrect, dropout):
+        super(RNNModel, self).__init__()
+
+        if num_layers <= 1:
+            dropout = 0.0
+        
+        self.rnn = nn.GRU(input_size=input_size, hidden_size=hidden_size, num_layers=num_layers, 
+            batch_first=True, dropout=dropout, bidirectional=bidrect)
+
+        self.number = (2 if bidrect else 1) * num_layers
+
+    def forward(self, x, mask):
+        '''
+            mask: (batch_size, seq_len) 
+            x: (batch_size, seq_len, input_size)
+        '''
+        lens = (mask).long().sum(dim=1)
+        lens, idx_sort = tc.sort(lens, descending=True)
+        _, idx_unsort = tc.sort(idx_sort)
+
+        x = x[idx_sort]
+
+        x = nn.utils.rnn.pack_padded_sequence(x, lens, batch_first=True)
+        self.rnn.flatten_parameters()
+        y, h = self.rnn(x)
+        y, lens = nn.utils.rnn.pad_packed_sequence(y, batch_first=True)
+
+        h = h.transpose(0,1).contiguous()   #make batch size first
+
+        y = y[idx_unsort]                   #(batch_size, seq_len, bid * hid_size)
+        h = h[idx_unsort]                   #(batch_size, number, hid_size)
+
+        return y, h
+
+class Contexualizer(nn.Module):
+    def __init__(self, input_size, hidden_size, num_layers=1, dropout=0.3):
+        super(Contexualizer, self).__init__()
+
+        self.rnn = RNNModel(input_size, hidden_size, num_layers, True, dropout)
+        self.output_size = hidden_size * 2
+
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        weights = self.rnn.rnn.all_weights
+        for w1 in weights:
+            for w2 in w1:
+                if len(list(w2.size())) <= 1:
+                    w2.data.fill_(0)
+                else: nn.init.xavier_normal_(w2.data, gain=1.414)
+
+    def forward(self, s, mask):
+        y = self.rnn(s, mask)[0]            # (batch_size, seq_len, 2 * hidden_size)
+
+        return y
+
+class ConcatAttention_Param(nn.Module):
+    def __init__(self, input_size, hidden_size, dropout=0.2):
+        super(ConcatAttention_Param, self).__init__()
+        self.ln = nn.Linear(input_size + hidden_size, hidden_size)
+        self.v = nn.Linear(hidden_size, 1, bias=False)
+        self.vq = nn.Parameter(tc.rand(hidden_size))
+        self.drop = nn.Dropout(dropout)
+
+        self.output_size = input_size
+        
+        self.reset_parameters()
+
+    def reset_parameters(self):
+
+        nn.init.xavier_uniform_(self.v.weight.data)
+        nn.init.xavier_uniform_(self.ln.weight.data)
+        self.ln.bias.data.fill_(0)
+
+    def forward(self, h, mask):
+        '''
+            h: (batch_size, len, input_size)
+            mask: (batch_size, len)
+        '''
+
+        vq = self.vq.view(1,1,-1).expand(h.size(0), h.size(1), self.vq.size(0))
+
+        s = self.v(tc.tanh(self.ln(tc.cat([h,vq],-1)))).squeeze(-1)    # (batch_size, len)
+        
+        s = s - ((mask == 0).float() * 10000)
+        a = tc.softmax(s, dim=1)
+
+        r = a.unsqueeze(-1) * h       # (batch_size, len, input_size)
+        r = tc.sum(r, dim=1)          # (batch_size, input_size)
+
+        return self.drop(r)
+
+ 
+def get_2dmask(mask_hq, mask_hp, siz=None):
+
+    if siz is None:
+        siz = (mask_hq.size(0), mask_hq.size(1), mask_hp.size(1))
+
+    mask_mat = 1
+    if mask_hq is not None:
+        mask_mat = mask_mat * mask_hq.unsqueeze(2).expand(siz)
+    if mask_hp is not None:
+        mask_mat = mask_mat * mask_hp.unsqueeze(1).expand(siz)
+    return mask_mat
+
+def Attention(hq, hp, mask_hq, mask_hp, my_method):
+    standard_size = (hq.size(0), hq.size(1), hp.size(1), hq.size(-1))
+    mask_mat = get_2dmask(mask_hq, mask_hp, standard_size[:-1])
+
+    hq_mat = hq.unsqueeze(2).expand(standard_size)
+    hp_mat = hp.unsqueeze(1).expand(standard_size)
+
+    s = my_method(hq_mat, hp_mat)           # (batch_size, len_q, len_p)
+
+    s = s - ((mask_mat == 0).float() * 10000)
+    a = tc.softmax(s, dim=1)
+
+    q = a.unsqueeze(-1) * hq_mat            #(batch_size, len_q, len_p, input_size)
+    q = tc.sum(q, dim=1)                    #(batch_size, len_p, input_size)
+
+    return q
+
+class ConcatAttention(nn.Module):
+    def __init__(self, input_size, hidden_size, dropout=0.2, input_size_2=-1):
+        super(ConcatAttention, self).__init__()
+
+        if input_size_2 < 0:
+            input_size_2 = input_size
+        self.ln = nn.Linear(input_size + input_size_2, hidden_size)
+        self.v = nn.Linear(hidden_size, 1, bias=False)
+        self.drop = nn.Dropout(dropout)
+
+        self.output_size = input_size
+
+        
+        self.reset_parameters()
+
+    def reset_parameters(self):
+
+        nn.init.xavier_uniform_(self.v.weight.data)
+        nn.init.xavier_uniform_(self.ln.weight.data)
+        self.ln.bias.data.fill_(0)
+
+    def my_method(self, hq_mat, hp_mat):
+        s = tc.cat([hq_mat, hp_mat], dim=-1)
+        s = self.v(tc.tanh(self.ln(s))).squeeze(-1)    #(batch_size, len_q, len_p)
+        return s
+
+    def forward(self, hq, hp, mask_hq=None, mask_hp=None):
+        '''
+            hq: (batch_size, len_q, input_size)
+            mask_hq: (batch_size, len_q)
+        '''
+        return self.drop(Attention(hq, hp, mask_hq, mask_hp, self.my_method))
+
+class MinusAttention(nn.Module):
+    def __init__(self, input_size, hidden_size, dropout=0.2):
+        super(MinusAttention, self).__init__()
+        self.ln = nn.Linear(input_size, hidden_size)
+        self.v = nn.Linear(hidden_size, 1, bias=False)
+
+        self.drop = nn.Dropout(dropout)
+        self.output_size = input_size
+        self.reset_parameters()
+
+    def reset_parameters(self):
+
+        nn.init.xavier_uniform_(self.v.weight.data)
+        nn.init.xavier_uniform_(self.ln.weight.data)
+        self.ln.bias.data.fill_(0)
+
+    def my_method(self, hq_mat, hp_mat):
+        s = hq_mat - hp_mat
+        s = self.v(tc.tanh(self.ln(s))).squeeze(-1)    #(batch_size, len_q, len_p) s[j,t]
+        return s
+
+    def forward(self, hq, hp, mask_hq=None, mask_hp=None):
+        return self.drop(Attention(hq, hp, mask_hq, mask_hp, self.my_method))
+
+class DotProductAttention(nn.Module):
+    def __init__(self, input_size, hidden_size, dropout=0.2):
+        super(DotProductAttention, self).__init__()
+        self.ln = nn.Linear(input_size, hidden_size)
+        self.v = nn.Linear(hidden_size, 1, bias=False)
+
+        self.drop = nn.Dropout(dropout)
+        self.output_size = input_size
+        self.reset_parameters()
+
+    def reset_parameters(self):
+
+        nn.init.xavier_uniform_(self.v.weight.data)
+        nn.init.xavier_uniform_(self.ln.weight.data)
+        self.ln.bias.data.fill_(0)
+
+    def my_method(self, hq_mat, hp_mat):
+        s = hq_mat * hp_mat
+        s = self.v(tc.tanh(self.ln(s))).squeeze(-1)    #(batch_size, len_q, len_p) s[j,t]
+        return s
+
+    def forward(self, hq, hp, mask_hq=None, mask_hp=None):
+        return self.drop(Attention(hq, hp, mask_hq, mask_hp, self.my_method))
+
+class BiLinearAttention(nn.Module):
+    def __init__(self, input_size, hidden_size, dropout=0.2, input_size_2=-1):
+        super(BiLinearAttention, self).__init__()
+
+        input_size_2 = input_size if input_size_2 < 0 else input_size_2
+
+        self.ln = nn.Linear(input_size_2, input_size)
+        self.drop = nn.Dropout(dropout)
+        self.output_size = input_size
+        
+        self.reset_parameters()
+
+    def reset_parameters(self):
+
+        nn.init.xavier_uniform_(self.ln.weight.data)
+        self.ln.bias.data.fill_(0)
+        
+    def my_method(self, hq, hp, mask_p):
+        # (bs, len, input_size)
+
+        hp = self.ln(hp)
+        hp = hp * mask_p.unsqueeze(-1)
+        s = tc.matmul(hq, hp.transpose(-1,-2))
+
+        return s
+
+    def forward(self, hq, hp, mask_hq=None, mask_hp=None):
+        standard_size = (hq.size(0), hq.size(1), hp.size(1), hq.size(-1))
+        mask_mat = get_2dmask(mask_hq, mask_hp, standard_size[:-1])
+
+        s = self.my_method(hq, hp, mask_hp)         # (batch_size, len_q, len_p)
+
+        s = s - ((mask_mat == 0).float() * 10000)
+        a = tc.softmax(s, dim=1)
+
+        hq_mat = hq.unsqueeze(2).expand(standard_size)
+        q = a.unsqueeze(-1) * hq_mat                #(batch_size, len_q, len_p, input_size)
+        q = tc.sum(q, dim=1)                      #(batch_size, len_p, input_size)
+
+        return self.drop(q)
+
+
+class AggAttention(nn.Module):
+    def __init__(self, input_size, hidden_size, dropout=0.2):
+        super(AggAttention, self).__init__()
+        self.ln = nn.Linear(input_size + hidden_size, hidden_size)
+        self.v = nn.Linear(hidden_size, 1, bias=False)
+        self.vq = nn.Parameter(tc.rand(hidden_size, 1))
+        self.drop = nn.Dropout(dropout)
+
+        self.output_size = input_size
+        
+        self.reset_parameters()
+
+    def reset_parameters(self):
+
+        nn.init.xavier_uniform_(self.vq.data)
+        nn.init.xavier_uniform_(self.v.weight.data)
+        nn.init.xavier_uniform_(self.ln.weight.data)
+        self.ln.bias.data.fill_(0)
+        self.vq.data = self.vq.data[:,0]
+
+
+    def forward(self, hs, mask):
+        '''
+            hs: [(batch_size, len_q, input_size), ...]
+            mask: (batch_size, len_q)
+        '''
+        
+        hs = tc.cat([h.unsqueeze(0) for h in hs], dim=0)# (4, batch_size, len_q, input_size)
+
+        vq = self.vq.view(1,1,1,-1).expand(hs.size(0), hs.size(1), hs.size(2), self.vq.size(0))
+
+        s = self.v(tc.tanh(self.ln(tc.cat([hs,vq],-1)))).squeeze(-1)# (4, batch_size, len_q)
+
+        s = s - ((mask.unsqueeze(0) == 0).float() * 10000)
+        a = tc.softmax(s, dim=0)
+
+        x = a.unsqueeze(-1) * hs
+        x = tc.sum(x, dim=0)#(batch_size, len_q, input_size)
+
+        return self.drop(x)
+
+class Aggragator(nn.Module):
+    def __init__(self, input_size, hidden_size, dropout=0.3):
+        super(Aggragator, self).__init__()
+
+        now_size = input_size
+        self.ln = nn.Linear(2 * input_size, 2 * input_size)
+
+        now_size = 2 * input_size
+        self.rnn = Contexualizer(now_size, hidden_size, 2, dropout)
+
+        now_size = self.rnn.output_size
+        self.agg_att = AggAttention(now_size, now_size, dropout)
+
+        now_size = self.agg_att.output_size
+        self.agg_rnn = Contexualizer(now_size, hidden_size, 2, dropout)
+
+        self.drop = nn.Dropout(dropout)
+
+        self.output_size = self.agg_rnn.output_size
+
+    def forward(self, qs, hp, mask):
+        '''
+            qs: [ (batch_size, len_p, input_size), ...]
+            hp: (batch_size, len_p, input_size)
+            mask if the same of hp's mask
+        '''
+
+        hs = [0 for _ in range(len(qs))]
+
+        for i in range(len(qs)):
+            q = qs[i]
+            x = tc.cat([q, hp], dim=-1)
+            g = tc.sigmoid(self.ln(x))
+            x_star = x * g
+            h = self.rnn(x_star, mask)
+
+            hs[i] = h
+
+        x = self.agg_att(hs, mask)      #(batch_size, len_p, output_size)
+        h = self.agg_rnn(x, mask)       #(batch_size, len_p, output_size)
+        return self.drop(h)
+
+
+class Mwan_Imm(nn.Module):
+    def __init__(self, input_size, hidden_size, num_class=3, dropout=0.2, use_allennlp=False):
+        super(Mwan_Imm, self).__init__()
+
+        now_size = input_size
+        self.enc_s1 = Contexualizer(now_size, hidden_size, 2, dropout)
+        self.enc_s2 = Contexualizer(now_size, hidden_size, 2, dropout)
+
+        now_size = self.enc_s1.output_size
+        self.att_c = ConcatAttention(now_size, hidden_size, dropout)
+        self.att_b = BiLinearAttention(now_size, hidden_size, dropout)
+        self.att_d = DotProductAttention(now_size, hidden_size, dropout)
+        self.att_m = MinusAttention(now_size, hidden_size, dropout)
+
+        now_size = self.att_c.output_size
+        self.agg = Aggragator(now_size, hidden_size, dropout)
+
+        now_size = self.enc_s1.output_size
+        self.pred_1 = ConcatAttention_Param(now_size, hidden_size, dropout)
+        now_size = self.agg.output_size
+        self.pred_2 = ConcatAttention(now_size, hidden_size, dropout, 
+                                                            input_size_2=self.pred_1.output_size)
+
+        now_size = self.pred_2.output_size
+        self.ln1 = nn.Linear(now_size, hidden_size)
+        self.ln2 = nn.Linear(hidden_size, num_class)
+
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        nn.init.xavier_uniform_(self.ln1.weight.data)
+        nn.init.xavier_uniform_(self.ln2.weight.data)
+        self.ln1.bias.data.fill_(0)
+        self.ln2.bias.data.fill_(0)
+
+    def forward(self, s1, s2, mas_s1, mas_s2):
+        hq = self.enc_s1(s1, mas_s1)                #(batch_size, len_q, output_size)
+        hp = self.enc_s1(s2, mas_s2)
+
+        mas_s1 = mas_s1[:,:hq.size(1)]
+        mas_s2 = mas_s2[:,:hp.size(1)]
+        mas_q, mas_p = mas_s1, mas_s2
+
+        qc = self.att_c(hq, hp, mas_s1, mas_s2)     #(batch_size, len_p, output_size)
+        qb = self.att_b(hq, hp, mas_s1, mas_s2)
+        qd = self.att_d(hq, hp, mas_s1, mas_s2)
+        qm = self.att_m(hq, hp, mas_s1, mas_s2)
+
+        ho = self.agg([qc,qb,qd,qm], hp, mas_s2)    #(batch_size, len_p, output_size)
+
+        rq = self.pred_1(hq, mas_q)                 #(batch_size, output_size)
+        rp = self.pred_2(ho, rq.unsqueeze(1), mas_p)#(batch_size, 1, output_size)
+        rp = rp.squeeze(1)                          #(batch_size, output_size)
+
+        rp = F.relu(self.ln1(rp))
+        rp = self.ln2(rp)
+
+        return rp
+
+class MwanModel(nn.Module):
+    def __init__(self, num_class, EmbLayer, args_of_imm={}, ElmoLayer=None):
+        super(MwanModel, self).__init__()
+
+        self.emb = EmbLayer
+
+        if ElmoLayer is not None:
+            self.elmo = ElmoLayer
+            self.elmo_preln = nn.Linear(3 * self.elmo.emb_size, self.elmo.emb_size)
+            self.elmo_ln = nn.Linear(args_of_imm["input_size"] + 
+                                                    self.elmo.emb_size, args_of_imm["input_size"])
+
+        else:
+            self.elmo = None
+
+
+        self.imm = Mwan_Imm(num_class=num_class, **args_of_imm)
+        self.drop = nn.Dropout(args_of_imm["dropout"])
+
+
+    def forward(self, words1, words2, str_s1=None, str_s2=None, *pargs, **kwargs):
+        '''
+            str_s is for elmo use , however we don't use elmo
+            str_s: (batch_size, seq_len, word_len)
+        '''
+
+        s1, s2 = words1, words2
+
+        mas_s1 = (s1 != 0).float()    # mas: (batch_size, seq_len)
+        mas_s2 = (s2 != 0).float()    # mas: (batch_size, seq_len)
+
+        mas_s1.requires_grad = False
+        mas_s2.requires_grad = False
+
+        s1_emb = self.emb(s1)
+        s2_emb = self.emb(s2)
+
+        if self.elmo is not None:
+            s1_elmo = self.elmo(str_s1)
+            s2_elmo = self.elmo(str_s2)
+
+            s1_elmo = tc.tanh(self.elmo_preln(tc.cat(s1_elmo, dim=-1)))
+            s2_elmo = tc.tanh(self.elmo_preln(tc.cat(s2_elmo, dim=-1)))
+
+            s1_emb = tc.cat([s1_emb, s1_elmo], dim=-1)
+            s2_emb = tc.cat([s2_emb, s2_elmo], dim=-1)
+
+            s1_emb = tc.tanh(self.elmo_ln(s1_emb))
+            s2_emb = tc.tanh(self.elmo_ln(s2_emb))
+
+        s1_emb = self.drop(s1_emb)
+        s2_emb = self.drop(s2_emb)
+
+        y = self.imm(s1_emb, s2_emb, mas_s1, mas_s2)
+
+        return {
+            Const.OUTPUT: y, 
+        }